APEX-I
MTP
Abliterix
GGUF
Qwen-AgentWorld-35B-A3B-MTP-Uncensored-APEX
English | Chinese README
AgentWorld world model · Abliterix uncensored · native MTP injected · APEX-I GGUFs
Important Warning
Qwen-AgentWorld is designed as an agentic environment simulation / language world model. It is intended to simulate agent-environment interactions and predict environment states from actions and history. It is not recommended as a general daily assistant, coding model, or autonomous Agent backbone unless you specifically need this world-model behavior.
About This Release
This package contains GGUF releases of Qwen-AgentWorld-35B-A3B, a Qwen3.5 MoE world model for agentic environment simulation.
The model was modified with an Abliterix uncensored LoRA, stream-merged into safetensors, injected with one native MTP layer from a compatible Qwen3.5-35B-A3B MTP shard, converted to BF16 GGUF, and quantized into APEX-I tiers using a model-specific imatrix.
The release keeps the BF16 GGUF, imatrix, and mmproj-F16.gguf for reproducibility and multimodal llama.cpp use.
Model Details
| Base model | Qwen/Qwen-AgentWorld-35B-A3B |
| Architecture | Qwen3.5 MoE |
| Parameters | 35B total, around 3B active per token |
| Experts | 256 routed experts, 8 active per token |
| Context | 262,144 tokens |
| MTP | 1 injected MTP layer, 785 tensors |
| License | Apache-2.0 |
Abliterix Build
The uncensored variant was produced with Abliterix. The selected search result was trial 54 out of 110.
| KL divergence | 0.0163 |
| Response length deviation | 0.04 std devs |
| Refusals | 21 / 100 |
The exported Abliterix LoRA was stream-merged into the full safetensors model before MTP injection and GGUF conversion.
MTP Extraction And Injection
The MTP layer was extracted from a compatible Qwen3.5-35B-A3B MTP shard and injected into the uncensored AgentWorld safetensors model.
| Tensor prefix | mtp.* |
| Tensor count | 785 |
| Injected shard | model-00022-of-00022.safetensors |
BenchLocal Results
| Model | TC | BF | HA | Ceiling | Practical | Retry |
|---|
| Original, thinking | 100 | 87 | 81 | 88.5 | 87.5 | 1 |
| Uncensored + MTP, thinking | 93 | 87 | 79 | 85.6 | 81.6 | 4 |
| Uncensored + MTP, no thinking | 83 | 83 | 63 | 75.0 | 75.0 | 0 |
| Qwen3.6-35B thinking baseline | 97 | 88 | 85 | 89.5 | 81.5 | 8 |
Thinking mode is strongly recommended for this model family.
Usage
llama.cpp server
hf download SC117/Qwen-AgentWorld-35B-A3B-MTP-Uncensored-GGUF --include "*.gguf" --local-dir ./models/qwen-agentworld-mtp-uncensored
./llama-server -m ./models/qwen-agentworld-mtp-uncensored/Qwen-AgentWorld-35B-A3B-MTP-Uncensored-APEX-I-Compact.gguf -ngl 99 -c 131072
vision + text
./llama-server -m ./models/qwen-agentworld-mtp-uncensored/Qwen-AgentWorld-35B-A3B-MTP-Uncensored-APEX-I-Compact.gguf --mmproj ./models/qwen-agentworld-mtp-uncensored/mmproj-F16.gguf -ngl 99 -c 131072
APEX Quantization Tiers
| File | Size | Profile | Best For |
|---|
*-APEX-I-Balanced.gguf | 24.14 GB | I-Balanced | Highest-retention all-rounder |
*-APEX-I-Quality.gguf | 21.68 GB | I-Quality | Quality-focused, IQ4_XS middle experts |
*-APEX-I-Compact.gguf | 15.75 GB | I-Compact | Recommended quality/size ratio |
*-APEX-I-Mini.gguf | 12.80 GB | I-Mini | Smallest practical tier |
The model-specific imatrix was generated from cleaned AgentWorld calibration text: PPL = 4.5880 +/- 0.01173.
Links
Citation
@misc{qwen-agentworld-2026,
title = {Qwen-AgentWorld: Language World Models for General Agents},
author = {Qwen Team},
year = {2026},
url = {https://github.com/QwenLM/Qwen-AgentWorld}
}
Install from pip and serve model
# Install vLLM from pip: pip install vllm# Start the vLLM server: vllm serve "SC117/Qwen-AgentWorld-35B-A3B-MTP-Uncensored-APEX-GGUF"# Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "SC117/Qwen-AgentWorld-35B-A3B-MTP-Uncensored-APEX-GGUF", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'